Building thermal wall mass provides a flexible heat capacity which can be effectively used for load shifting, thus enabling demand side management (DSM). The most crucial barrier for a practical application of buildings as short term heat storage is the lack of knowledge about the building physical properties. This study presents a model identification approach for forecasting the building thermal response based on grey-box models. The model parameters are estimated by optimizing the model output to historical data under the consideration of plausible physical constraints. The thermal flexibility of a building and therewith its demand side management potential can be determined based on the assessed model parameters. This study uses measurements from buildings in normal operation to evaluate the thermal prediction of grey-box models despite the stochastic events within the training data. In order to find the best level of model complexity, four grey-box models were compared in their ability to forecast the building indoor temperature behaviour. The analysis revealed that a two-capacity model structure with an additional consideration of the indoor air as a mass-less node (4R2C-model) enables the most accurate qualitative prediction of the indoor temperature. Further, the general validity of the model structure allows for its application on different buildings types. (C) 2016 Elsevier B.V. All rights reserved.